Research On Early Diagnosis Algorithm Of Alzheimer's Disease Based On Transfer Learning

Mild cognitive impairment(MCI)is an important transitional stage in the early stage of Alzheimer's disease(AD)dementia.Prevention and treatment of MCI stage plays an important role in delaying the onset of dementia.Neuroimaging technology is a widely used research method in the early diagnosis of AD brain diseases.However,AD is concealed and slow,sample collection is difficult and the sample size is small.Most of the training set samples used in machine learning-based AD diagnostic studies come from the same field of study.In order to make full use of the prior knowledge of related fields and solve the problem of insufficient sample size,transfer learning has been proposed and can solve such problems better.Therefore,based on magnetic resonance imaging(MRI)data,combined with prior information in related fields,this thesis proposes an early diagnosis algorithm for AD based on transfer learning,which realizes the classification of MCI patients.The specific work mainly includes:Firstly,an early diagnosis algorithm for AD based on discriminative transfer feature learning is proposed.The algorithm uses the data of the target domain(i.e.,MCI)and the source domain(i.e.,AD and normal control group(NC))to obtain the most discriminative feature subset for MCI classification prediction.Specifically,firstly,according to the correlation between the target domain and the source domain data,the transfer component analysis method based on the maximum mean difference(MMD)is used to narrow the data distribution difference between the source domain and the target domain,and obtain an effective feature subset.Secondly,Considering that the TCA algorithm ignores the correlation between sample categories and sample features in each domain,which may make the original similar data produce a large deviation after projection,a discriminant optimization term based on the intra-class divergence matrix is added.Discriminate the optimization term to better utilize the discriminative information between samples;finally,use the support vector machine(SVM)method to classify and predict MCI patients through the obtained discriminative feature subsets.The experimental results show that the method has a good classification performance in the classification of MCIc and MCInc patients,the classification accuracy is 75.13%,and the AUC value is 0.7857.Secondly,an early diagnosis algorithm for AD based on ensemble transfer learning is proposed.The purpose is to make full use of the discriminant information of the sample to realize the transfer of knowledge between domains and improve the classification performance of the classification model.The basic ideas are as follows: Firstly,based on the problem that the traditional feature transfer TCA algorithm is insufficiently utilized for sample structure information and tag information,a feature transfer learning algorithm based on joint distributionadaptive(JDA)is adopted to narrow the edge distribution difference between domains.Differences in conditional distribution,extracting shared features between domains,and realizing the transfer of knowledge between domains;secondly,in order to obtain more accurate AD classification effects,it is proposed to use GBDT,XGBoost and Adaboost as the base classifier for ensemble learning,and use feature subset training.The base classifier is obtained,and the final classification result is obtained by the weighted voting method of multiple base classifiers,and the early diagnosis of AD is realized.The experimental results show that the method has a good classification prediction effect on the ADNI standard data set,and its classification accuracy is 78.12%,and the AUC value is 0.8257.